ELYZA-Thinking-1.0-Qwen-32B Overview

ELYZA-Thinking-1.0-Qwen-32B is a 32.8 billion parameter language model developed by ELYZA, Inc., specifically engineered to excel in Japanese reasoning tasks. It is built upon the robust Qwen2.5-32B-Instruct foundation, with a focus on enhancing its analytical and problem-solving abilities through specialized post-training.

Key Capabilities & Training

• Enhanced Japanese Reasoning: The model's primary differentiator is its significantly improved reasoning performance in Japanese, achieved through a unique post-training methodology.
• Imitation Learning with CoT: Training involved imitation learning on synthetic data, which incorporated extensive Chains of Thought (CoT). These CoT sequences were generated using an advanced Monte Carlo Tree Search (MCTS)-based algorithm, allowing the model to learn complex reasoning patterns.
• Qwen Architecture: Leverages the Qwen2.5-32B-Instruct base, providing a strong general-purpose language understanding foundation.

Usage Recommendations

• Inference: The model is readily usable with the Hugging Face Transformers library. Example code is provided for quick integration.
• Deployment: For high-performance deployment, vLLM is recommended, supporting an OpenAI-Compatible Server setup.
• Parameter Tuning: For optimal results and to prevent repetitive outputs, ELYZA recommends setting temperature between 0.5 and 0.7 and top_p to 0.95 during generation.